On this page
An idea can be understood perfectly well by the person who has it and still be difficult to implement. In software development, that gap is familiar. People describe what they need using the language of their work. An implementation needs something more precise: responsibilities, constraints, interactions and a way to decide whether the result is acceptable.
Working with an AI assistant changes the speed at which those two descriptions can meet. It does not make them equivalent. I have found that the useful part of the conversation is not how quickly a proposed solution appears. It is whether the intended outcome can be made sufficiently clear to examine against a real system.
An outcome is not yet a specification
"Make the research easier to navigate" conveys a direction but leaves several decisions open. What must remain visible? Which information belongs on an article page, and which belongs in its archive? How should the page behave on a phone? What happens to navigation when the article contains nine sections rather than three?
Those questions cannot all be answered by making the initial description longer. Some become apparent only when there is a working page to inspect. The aim is therefore to establish enough shared understanding to make a bounded change, while keeping room to learn from the outcome.
The same difficulty appears in conversational analytics. A question such as "Which customer bought the most?" sounds clear until most needs to become an analytical operation: transaction count, quantity or spending, over which period, with what treatment of returns? In determinIQ, the distinction is handled by a governed query contract. Software engineering has different authority and verification requirements, but the underlying lesson is related: a plausible interpretation is not automatically the intended one.
Find the boundary before changing the implementation
With this website, we established early that the theme owns presentation and that the separate John Stutely Core must-use plugin owns the editorial structures, metadata and related application behaviour. That separation gives an otherwise broad instruction a practical boundary.
A request to adjust article colours belongs in the theme. A request to add research drafts belongs in the Core publishing workflow. A change that touches both requires an explicit account of which component is responsible for each part. Without that distinction, a quick solution can create duplicated behaviour that becomes difficult to maintain.
This is where I see a substantial role for an AI assistant: examining an existing implementation, explaining which components are involved, proposing a limited change and making its assumptions available for challenge. The proposed implementation is useful evidence of an interpretation, not proof that the interpretation was right.
Turn an intention into an observable result
I find it helpful to distinguish an intended outcome from what we can actually inspect. For a research article, that might mean preserving its text while introducing a diagram, confirming the diagram does not overflow at mobile widths, and ensuring that light, dark and high-contrast presentations remain readable. For an importer, it might mean creating only missing drafts, avoiding duplicate records on a second run, and never replacing an article somebody has edited.
Those are materially different requirements from "add a diagram" or "import the content". They define what the change must respect as well as what it must accomplish.
Research figure 01
From intention to something testable
A conceptual path, not a literal sequence of prompts
- PURPOSEDescribe the needed outcomeWhat should be possible, useful or clearer?
- EXISTING SYSTEMEstablish the boundariesInspect responsibilities, constraints and dependencies
- IMPLEMENTABLE CHANGEMake the work specificChoose a bounded intervention and its expected behaviour
- ACCEPTANCEInspect what happenedRendered output, actual behaviour, tests and regressions
- JUDGEMENTAccept, adjust or reconsiderThe next decision follows the observed result
Human review may revisit any stage when assumptions change.
A brief example: publishing research drafts
The requirement for this website was not simply to move text from one place to another. The publications needed their correct editorial collection, topics, evidence references and relationships, with no manual assembly. They also needed to remain drafts until review and approval.
An importer could satisfy the first part and still be unsuitable. If running it twice produced duplicates, or if a subsequent update overwrote edits, the convenience would come at an unacceptable cost. We therefore made stable identities, idempotent imports and preservation of editorial state part of the acceptance criteria. The same requirements had to survive incremental updates to the importer.
There was a practical reminder of why the installation context matters too. At one point an incomplete must-use plugin installation left the site without its editorial menus. The remedy was not another feature; it was restoring the correct directory structure and packaging complete dependencies. A syntactically sound module is of little use if the host application cannot load it.
What remains human work
None of this removes the need to decide which requirement is worth meeting, where a boundary belongs or whether a compromise is acceptable. Those are not merely translation tasks. They are decisions about the product and its users.
It is also important not to confuse a successful local implementation with a general-purpose method that has been experimentally validated. The observations here come from work on particular applications, with particular constraints and a human reviewer involved throughout. They show how the approach can operate in practice, not a quantified improvement over every alternative.
The further question for this research is how much of the translation can be made quicker and more reliable while keeping the requirements understandable, contestable and grounded in the actual implementation. The objective is not to produce longer instructions. It is to reduce the distance between what somebody means and what can responsibly be built.
Research context
Research programme: From intention to implementation: human-directed AI-assisted engineering
Topic: AI-assisted engineering
Explore further
Further reading is listed below.